Hierarchical Semantic Labeling with Adaptive Confidence

Jim Davis, Tong Liang, James Enouen, Roman Ilin

Abstract: In real-world applications with semantic classification labels (‘dog’, ‘car’, ‘chair’, etc.), it would be advantageous to identify any unconfident classification and then determine if a less specific label could instead be reliably established. In this work, we present a hierarchical estimation and inference approach using a semantic concept tree to provide an appropriate generalized label when needed. The proposed method has several advantages, including the ability to work with any logit/softmax-based semantic label classifier, the ability to correct many misclassified labels while not introducing any new errors, and a statistical guarantee of confidence for the final labels. We additionally provide a new set of hierarchical metrics to properly evaluate the approach. Multiple synthetic and real datasets are examined to demonstrate how the framework can quickly and efficiently resolve unconfident predictions.